Tidy Up Date/Time Ranges (neatRanges) Collapse, partition, combine, fill gaps in and expand date/time ranges.
Selecting Matched Samples from Multiply Imputed Datasets (MatchIt.mice) Selects matched samples from the control and treatment groups of each imputed datasets and estimates the weight of each individual in the complete dist …
Example Datasets for a Learning Guide to R (lgrdata) A largish collection of example datasets, including several classics. Many of these datasets are well suited for regression, classification, and visual …
Simple Git Client (gert) Minimal git client for R based on ‘libgit2’ <https://…/>. This package requires a somewha …
There has been much discussion recently about how fairness should be measured or enforced in classification. Individual Fairness [Dwork, Hardt, Pitassi, Reingold, Zemel, 2012], which requires that similar individuals be treated similarly, is a highly appealing definition as it gives strong guarantees on treatment of individuals. Unfortunately, the need for a task-specific similarity metric has prevented its use in practice. In this work, we propose a solution to the problem of approximating a metric for Individual Fairness based on human judgments. Our model assumes that we have access to a human fairness arbiter, who can answer a limited set of queries concerning similarity of individuals for a particular task, is free of explicit biases and possesses sufficient domain knowledge to evaluate similarity. Our contributions include definitions for metric approximation relevant for Individual Fairness, constructions for approximations from a limited number of realistic queries to the arbiter on a sample of individuals, and learning procedures to construct hypotheses for metric approximations which generalize to unseen samples under certain assumptions of learnability of distance threshold functions.
We address the challenge of designing optimal adversarial noise algorithms for settings where a learner has access to multiple classifiers. We demonstrate how this problem can be framed as finding strategies at equilibrium in a two-player, zero-sum game between a learner and an adversary. In doing so, we illustrate the need for randomization in adversarial attacks. In order to compute Nash equilibrium, our main technical focus is on the design of best response oracles that can then be implemented within a Multiplicative Weights Update framework to boost deterministic perturbations against a set of models into optimal mixed strategies. We demonstrate the practical effectiveness of our approach on a series of image classification tasks using both linear classifiers and deep neural networks.
Stochastic Volatility (SV) models are widely used in the financial sector while Long Short-Term Memory (LSTM) models have been successfully used in many large-scale industrial applications of Deep Learning. Our article combines these two methods non trivially and proposes a model for capturing the dynamics of financial volatility process, which we call the LSTM-SV model. The proposed model overcomes the short-term memory problem in conventional SV models, is able to capture non-linear dependence in the latent volatility process, and often has a better out-of-sample forecast performance than SV models. The conclusions are illustrated through simulation studies and applications to three financial time series datasets: US stock market weekly index SP500, Australian stock weekly index ASX200 and Australian-US dollar daily exchange rates. We argue that there are significant differences in the underlying dynamics between the volatility process of SP500 and ASX200 datasets and that of the exchange rate dataset. For the stock index data, there is strong evidence of long-term memory and non-linear dependence in the volatility process, while this is not the case for the exchange rates. An user-friendly software package together with the examples reported in the paper are available at https://…/vbayeslab.
Semantic segmentation and instance level segmentation made substantial progress in recent years due to the emergence of deep neural networks (DNNs). A number of deep architectures with Convolution Neural Networks (CNNs) were proposed that surpass the traditional machine learning approaches for segmentation by a large margin. These architectures predict the directly observable semantic category of each pixel by usually optimizing a cross entropy loss. In this work we push the limit of semantic segmentation towards predicting semantic labels of directly visible as well as occluded objects or objects parts, where the network’s input is a single depth image. We group the semantic categories into one background and multiple foreground object groups, and we propose a modification of the standard cross-entropy loss to cope with the settings. In our experiments we demonstrate that a CNN trained by minimizing the proposed loss is able to predict semantic categories for visible and occluded object parts without requiring to increase the network size (compared to a standard segmentation task). The results are validated on a newly generated dataset (augmented from SUNCG) dataset.
Recurrent neural networks (RNNs) are commonly applied to clinical time-series data with the goal of learning patient risk stratification models. Their effectiveness is due, in part, to their use of parameter sharing over time (i.e., cells are repeated hence the name recurrent). We hypothesize, however, that this trait also contributes to the increased difficulty such models have with learning relationships that change over time. Conditional shift, i.e., changes in the relationship between the input X and the output y, arises if the risk factors for the event of interest change over the course of a patient admission. While in theory, RNNs and gated RNNs (e.g., LSTMs) in particular should be capable of learning time-varying relationships, when training data are limited, such models often fail to accurately capture these dynamics. We illustrate the advantages and disadvantages of complete weight sharing (RNNs) by comparing an LSTM with shared parameters to a sequential architecture with time-varying parameters on three clinically-relevant prediction tasks: acute respiratory failure (ARF), shock, and in-hospital mortality. In experiments using synthetic data, we demonstrate how weight sharing in LSTMs leads to worse performance in the presence of conditional shift. To improve upon the dichotomy between complete weight sharing vs. no weight sharing, we propose a novel RNN formulation based on a mixture model in which we relax weight sharing over time. The proposed method outperforms standard LSTMs and other state-of-the-art baselines across all tasks. In settings with limited data, relaxed weight sharing can lead to improved patient risk stratification performance.
The I.I.D. hypothesis between training data and testing data is the basis of a large number of image classification methods. Such a property can hardly be guaranteed in practical cases where the Non-IIDness is common, leading to instable performances of these models. In literature, however, the Non-I.I.D. image classification problem is largely understudied. A key reason is the lacking of a well-designed dataset to support related research. In this paper, we construct and release a Non-I.I.D. image dataset called NICO, which makes use of contexts to create Non-IIDness consciously. Extended experimental results and anslyses demonstrate that the NICO dataset can well support the training of a ConvNet model from scratch, and NICO can support various Non-I.I.D. situations with sufficient flexibility compared to other datasets.
Human learners have the natural ability to use knowledge gained in one setting for learning in a different but related setting. This ability to transfer knowledge from one task to another is essential for effective learning. In this paper, we study transfer learning in the context of nonparametric classification based on observations from different distributions under the posterior drift model, which is a general framework and arises in many practical problems. We first establish the minimax rate of convergence and construct a rate-optimal two-sample weighted -NN classifier. The results characterize precisely the contribution of the observations from the source distribution to the classification task under the target distribution. A data-driven adaptive classifier is then proposed and is shown to simultaneously attain within a logarithmic factor of the optimal rate over a large collection of parameter spaces. Simulation studies and real data applications are carried out where the numerical results further illustrate the theoretical analysis. Extensions to the case of multiple source distributions are also considered.
We propose AutoGrow to automate depth discovery in Deep Neural Networks (DNNs): starting from a shallow seed architecture, AutoGrow grows new layers if the growth improves the accuracy; otherwise, the growth stops and the network depth is discovered. The residual and plain blocks are used as growing sub-modules to study DNNs with and without shortcuts. We propose generic growing and stopping policies to minimize human efforts spent on the optimal depth search. Our experiments show that by applying the same policy to different tasks, AutoGrow can always discover network depth effectively and achieve state-of-the-art accuracy on various datasets of MNIST, FashionMNIST, SVHN, CIFAR10, CIFAR100 and ImageNet. Comparing to Neural Architecture Search (NAS) that often designs a gigantic search space and consumes tremendous resources, AutoGrow lies at the other end of the research spectrum: it focuses on efficient depth discovery and reduces the growing and searching time to a level similar to that of training a single DNN. Thus, AutoGrow is able to scale up to large datasets such as ImageNet. Our study also reveals that previous Network Morphism is sub-optimal for increasing layer depth. Finally, we demonstrate that AutoGrow enables the training of deeper plain networks, which has been problematic even using Batch Normalization.
We propose algorithms with state-of-the-art \emph{dynamic regret} bounds for un-discounted reinforcement learning under drifting non-stationarity, where both the reward functions and state transition distributions are allowed to evolve over time. Our main contributions are: 1) A tuned Sliding Window Upper-Confidence bound for Reinforcement Learning with Confidence-Widening (\texttt{SWUCRL2-CW}) algorithm, which attains low dynamic regret bounds against the optimal non-stationary policy in various cases. 2) The Bandit-over-Reinforcement Learning (\texttt{BORL}) framework that further permits us to enjoy these dynamic regret bounds in a parameter-free manner.
Interactive NLP is a promising paradigm to close the gap between automatic NLP systems and the human upper bound. Preference-based interactive learning has been successfully applied, but the existing methods require several thousand interaction rounds even in simulations with perfect user feedback. In this paper, we study preference-based interactive summarisation. To reduce the number of interaction rounds, we propose the Active Preference-based ReInforcement Learning (APRIL) framework. APRIL uses Active Learning to query the user, Preference Learning to learn a summary ranking function from the preferences, and neural Reinforcement Learning to efficiently search for the (near-)optimal summary. Our results show that users can easily provide reliable preferences over summaries and that APRIL outperforms the state-of-the-art preference-based interactive method in both simulation and real-user experiments.
A novel resampling framework is proposed to evaluate the robustness and generalization capability of deep learning models with respect to distribution shift. We use Auto Encoder Variational Bayes to find a latent representation of the data, on which a Variational Gaussian Mixture Model is applied to deliberately create distribution shift by dividing the dataset into different clusters. Wasserstein distance is used to characterize the extent of distribution shift between the training and the testing data splits. We compare several conventional Convolutional Neural Network (CNN) architectures as well as Bayesian CNN models for image classification on the Fashion-MNIST dataset to assess their robustness under the deliberately created distribution shift.
Deep Neural Networks have shown tremendous success in the area of object recognition, image classification and natural language processing. However, designing optimal Neural Network architectures that can learn and output arbitrary graphs is an ongoing research problem. The objective of this survey is to summarize and discuss the latest advances in methods to Learn Representations of Graph Data. We start by identifying commonly used types of graph data and review basics of graph theory. This is followed by a discussion of the relationships between graph kernel methods and neural networks. Next we identify the major approaches used for learning representations of graph data namely: Kernel approaches, Convolutional approaches, Graph neural networks approaches, Graph embedding approaches and Probabilistic approaches. A variety of methods under each of the approaches are discussed and the survey is concluded with a brief discussion of the future of learning representation of graph data.
We consider a nonparametric heteroscedastic time series regression model and suggest testing procedures to detect changes in the conditional variance function. The tests are based on a sequential marked empirical process and thus combine classical CUSUM tests with marked empirical process approaches known from goodness-of-fit testing. The tests are consistent against general alternatives of a change in the conditional variance function, a feature that classical CUSUM tests are lacking. We derive a simple limiting distribution and in the case of univariate covariates even obtain asymptotically distribution-free tests. We demonstrate the good performance of the tests in a simulation study and consider exchange rate data as a real data application.
Online change-point detection (OCPD) is important for application in various areas such as finance, biology, and the Internet of Things (IoT). However, OCPD faces major challenges due to high-dimensionality, and it is still rarely studied in literature. In this paper, we propose a novel, online, graph-based, change-point detection algorithm to detect change of distribution in low- to high-dimensional data. We introduce a similarity measure, which is derived from the graph-spanning ratio, to test statistically if a change occurs. Through numerical study using artificial online datasets, our data-driven approach demonstrates high detection power for high-dimensional data, while the false alarm rate (type I error) is controlled at a nominal significant level. In particular, our graph-spanning approach has desirable power with small and multiple scanning window, which allows timely detection of change-point in the online setting.
Probabilistic programming has emerged as a powerful paradigm in statistics, applied science, and machine learning: by decoupling modelling from inference, it promises to allow modellers to directly reason about the processes generating data. However, the performance of inference algorithms can be dramatically affected by the parameterisation used to express a model, requiring users to transform their programs in non-intuitive ways. We argue for automating these transformations, and demonstrate that mechanisms available in recent modeling frameworks can implement non-centring and related reparameterisations. This enables new inference algorithms, and we propose two: a simple approach using interleaved sampling and a novel variational formulation that searches over a continuous space of parameterisations. We show that these approaches enable robust inference across a range of models, and can yield more efficient samplers than the best fixed parameterisation.
Human behavior expression and experience are inherently multi-modal, and characterized by vast individual and contextual heterogeneity. To achieve meaningful human-computer and human-robot interactions, multi-modal models of the users states (e.g., engagement) are therefore needed. Most of the existing works that try to build classifiers for the users states assume that the data to train the models are fully labeled. Nevertheless, data labeling is costly and tedious, and also prone to subjective interpretations by the human coders. This is even more pronounced when the data are multi-modal (e.g., some users are more expressive with their facial expressions, some with their voice). Thus, building models that can accurately estimate the users states during an interaction is challenging. To tackle this, we propose a novel multi-modal active learning (AL) approach that uses the notion of deep reinforcement learning (RL) to find an optimal policy for active selection of the users data, needed to train the target (modality-specific) models. We investigate different strategies for multi-modal data fusion, and show that the proposed model-level fusion coupled with RL outperforms the feature-level and modality-specific models, and the naive AL strategies such as random sampling, and the standard heuristics such as uncertainty sampling. We show the benefits of this approach on the task of engagement estimation from real-world child-robot interactions during an autism therapy. Importantly, we show that the proposed multi-modal AL approach can be used to efficiently personalize the engagement classifiers to the target user using a small amount of actively selected users data.
When time series are organized into hierarchies, the forecasts have to satisfy some summing constraints. Forecasts which are independently generated for each time series (base forecasts) do not satisfy the constraints. Reconciliation algorithms adjust the base forecast in order to satisfy the summing constraints: in general they also improve the accuracy. We present a novel reconciliation algorithm based on Bayes’ rule; we discuss under which assumptions it is optimal and we show in extensive experiments that it compares favorably to the state-of-the-art reconciliation methods.
Typically, recommender systems from any domain, be it movies, music, restaurants, etc., are organized in a centralized fashion. The service provider holds all the data, biases in the recommender algorithms are not transparent to the user, and the service providers often create lock-in effects making it inconvenient for the user to switch providers. In this paper, we argue that the user’s smartphone already holds a lot of the data that feeds into typical recommender systems for movies, music, or POIs. With the ubiquity of the smartphone and other users in proximity in public places or public transportation, data can be exchanged directly between users in a device-to-device manner. This way, each smartphone can build its own database and calculate its own recommendations. One of the benefits of such a system is that it is not restricted to recommendations for just one user – ad-hoc group recommendations are also possible. While the infrastructure for such a platform already exists – the smartphones already in the palms of the users – there are challenges both with respect to the mobile recommender system platform as well as to its recommender algorithms. In this paper, we present a mobile architecture for the described system – consisting of data collection, data exchange, and recommender system – and highlight its challenges and opportunities.
Empirical evidence shows that ensembles, such as bagging, boosting, random and rotation forests, generally perform better in terms of their generalization error than individual classifiers. To explain this performance, Schapire et al. (1998) developed an upper bound on the generalization error of an ensemble based on the margins of the training data, from which it was concluded that larger margins should lead to lower generalization error, everything else being equal. Many other researchers have backed this assumption and presented tighter bounds on the generalization error based on either the margins or functions of the margins. For instance, Shen and Li (2010) provide evidence suggesting that the generalization error of a voting classifier might be reduced by increasing the mean and decreasing the variance of the margins. In this article we propose several techniques and empirically test whether the current state of research in explaining ensemble performance holds. We evaluate the proposed methods through experiments with real and simulated data sets.
In-memory computing is a promising non-von Neumann approach where certain computational tasks are performed within memory units by exploiting the physical attributes of memory devices. For instance, crossbar arrays of resistive memory devices can be used to store a matrix and perform analog matrix-vector multiplications at constant O(1) time complexity without intermediate movements of data. This functionality is very appealing for making energy-efficient deep learning inference hardware, where the weights of the neural network layers would be encoded in such crossbar arrays. However, due to device variability and noise, the network needs to be trained in a specific way so that transferring the digitally trained weights to the analog resistive memory devices will not result in appreciable loss of accuracy. Here, we introduce a methodology to train ResNet-type convolutional neural networks that results in almost no accuracy loss when transferring weights to analog in-memory computing hardware based on phase-change memory (PCM). Our experimental results demonstrate an as-programmed classification accuracy of 93.69% on the CIFAR-10 dataset with ResNet-32, which stays above 92.6% over a one day period, where each of the 361,722 synaptic weights of the network is programmed on just two PCM devices organized in a differential configuration.
In this work we show that Evolution Strategies (ES) are a viable method for learning non-differentiable parameters of large supervised models. ES are black-box optimization algorithms that estimate distributions of model parameters; however they have only been used for relatively small problems so far. We show that it is possible to scale ES to more complex tasks and models with millions of parameters. While using ES for differentiable parameters is computationally impractical (although possible), we show that a hybrid approach is practically feasible in the case where the model has both differentiable and non-differentiable parameters. In this approach we use standard gradient-based methods for learning differentiable weights, while using ES for learning non-differentiable parameters – in our case sparsity masks of the weights. This proposed method is surprisingly competitive, and when parallelized over multiple devices has only negligible training time overhead compared to training with gradient descent. Additionally, this method allows to train sparse models from the first training step, so they can be much larger than when using methods that require training dense models first. We present results and analysis of supervised feed-forward models (such as MNIST and CIFAR-10 classification), as well as recurrent models, such as SparseWaveRNN for text-to-speech.
Regression models describing the joint distribution of multivariate response variables conditional on covariate information have become an important aspect of contemporary regression analysis. However, a limitation of such models is that they often rely on rather simplistic assumptions, e.g. a constant dependency structure that is not allowed to vary with the covariates. We propose a general framework for multivariate conditional transformation models that overcomes such limitations and describes the full joint distribution in simple, interpretable terms. Among the particular merits of the framework are that it can be embedded into likelihood-based inference and allows the dependence structure to vary with the covariates. In addition, the framework scales beyond bivariate response situations, which were the main focus of most earlier investigations. We illustrate the application of multivariate conditional transformation models in a trivariate analysis of childhood undernutrition and demonstrate empirically that even complex multivariate data-generating processes can be inferred from observations.
Distributed learning and random projections are the most common techniques in large scale nonparametric statistical learning. In this paper, we study the generalization properties of kernel ridge regression using both distributed methods and random features. Theoretical analysis shows the combination remarkably reduces computational cost while preserving the optimal generalization accuracy under standard assumptions. In a benign case, partitions and random features are sufficient to achieve learning rate, where is the labeled sample size. Further, we derive more refined results by using additional unlabeled data to enlarge the number of partitions and by generating features in a data-dependent way to reduce the number of random features.
General purpose relation extractors, which can model arbitrary relations, are a core aspiration in information extraction. Efforts have been made to build general purpose extractors that represent relations with their surface forms, or which jointly embed surface forms with relations from an existing knowledge graph. However, both of these approaches are limited in their ability to generalize. In this paper, we build on extensions of Harris’ distributional hypothesis to relations, as well as recent advances in learning text representations (specifically, BERT), to build task agnostic relation representations solely from entity-linked text. We show that these representations significantly outperform previous work on exemplar based relation extraction (FewRel) even without using any of that task’s training data. We also show that models initialized with our task agnostic representations, and then tuned on supervised relation extraction datasets, significantly outperform the previous methods on SemEval 2010 Task 8, KBP37, and TACRED.
Capsule Networks attempt to represent patterns in images in a way that preserves hierarchical spatial relationships. Additionally, research has demonstrated that these techniques may be robust against adversarial perturbations. We present an improvement to training capsule networks with added robustness via non-parametric kernel methods. The representations learned through the capsule network are used to construct covariance kernels for Gaussian processes (GPs). We demonstrate that this approach achieves comparable prediction performance to Capsule Networks while improving robustness to adversarial perturbations and providing a meaningful measure of uncertainty that may aid in the detection of adversarial inputs.
Recent successes in visual recognition can be primarily attributed to feature representation, learning algorithms, and the ever-increasing size of labeled training data. Extensive research has been devoted to the first two, but much less attention has been paid to the third. Due to the high cost of manual labeling, the size of recent efforts such as ImageNet is still relatively small in respect to daily applications. In this work, we mainly focus on how to automatically generate identifying image data for a given visual concept on a vast scale. With the generated image data, we can train a robust recognition model for the given concept. We evaluate the proposed webly supervised approach on the benchmark Pascal VOC 2007 dataset and the results demonstrates the superiority of our proposed approach in image data collection.
As a new way to train generative models, generative adversarial networks (GANs) have achieved considerable success in image generation, and this framework has also recently been applied to data with graph structures. We identify the drawbacks of existing deep frameworks for generating graphs, and we propose labeled-graph generative adversarial networks (LGGAN) to train deep generative models for graph-structured data with node labels. We test the approach on various types of graph datasets, such as collections of citation networks and protein graphs. Experiment results show that our model can generate diverse labeled graphs that match the structural characteristics of the training data and outperforms all baselines in terms of quality, generality, and scalability. To further evaluate the quality of the generated graphs, we apply it to a downstream task for graph classification, and the results show that LGGAN can better capture the important aspects of the graph structure.
We propose an ensemble technique for converting any classifier into a computationally secure classifier. We define a simpler security problem for random binary classifiers and prove a reduction from this model to the security of the overall ensemble classifier. We provide experimental evidence of the security of our random binary classifiers, as well as empirical results of the adversarial accuracy of the overall ensemble to black-box attacks. Our construction crucially leverages hidden randomness in the multiclass-to-binary reduction.
The analysis of natural disaster-related multimedia content got great attention in recent years. Being one of the most important sources of information, social media have been crawled over the years to collect and analyze disaster-related multimedia content. Satellite imagery has also been widely explored for disasters analysis. In this paper, we survey the existing literature on disaster detection and analysis of the retrieved information from social media and satellites. Literature on disaster detection and analysis of related multimedia content on the basis of the nature of the content can be categorized into three groups, namely (i) disaster detection in text; (ii) analysis of disaster-related visual content from social media; and (iii) disaster detection in satellite imagery. We extensively review different approaches proposed in these three domains. Furthermore, we also review benchmarking datasets available for the evaluation of disaster detection frameworks. Moreover, we provide a detailed discussion on the insights obtained from the literature review, and identify future trends and challenges, which will provide an important starting point for the researchers in the field.Natural Disasters Detection in Social Media and Satellite imagery: a survey
Data science is reaching the end of its hype cycle and the skills for becoming a data scientist are changing. Besides of fiddling about the best performant machine learning model, it is more important than ever to make data science with business impact. In an Analytics & Data Science Meetup by DataIku in Berlin (Germany) I discussed the current skill set that help define and build up a data science workflow delivering real value. The slides of my talk you can find here.
LSTM Recurrent Neural Networks turn out to be a good choice for time series prediction task, however the algorithm relies on the assumption that we have sufficient training and testing data coming from the same distribution. The challenge is that time-series data usually exhibit time-varying characteristic, which may lead to a wide variability between old and new data. In this blog we want to test to which extent transfer learning and general domain training of models help tackle the above-mentioned problem.
Machine learning is being deployed to do large-scale decision making, which can strongly impact the life of individuals. By not considering and analysing such scenarios, we may end up building models that fail to treat societies equally and even infringe anti-discrimination laws. There are several algorithmic interventions to identify unfair treatment based on what is considered to be fair. In this article, we will visit these and explain their benefits and limitations with a case study.
In a recent class of Network Analytics, we were asked to visualise correlations between stocks. As an investor, you’re interested in diversifying risk by selecting different types of them. You might therefore want visualise which stocks behave similarly (positive correlations) or very differently (negative correlations). Using a dataset with the prices of selected stocks over time, we’ll create a correlation matrix that we’ll visualise with Networkx.
Optimising your R code is not always the priority. But when you run out of memory, or it just takes too long, you start to wonder if there are better ways to do things! In this blog post, I will show you my way of optimising my R code and the process behind it. As an example, I will use the code from my last blog post about gas prices in Germany – if you haven’t read it yet, click here! There are different steps to optimise your code. From debugging and profiling over benchmarking to rethinking the whole method. You have to repeat these steps until you are satisfied with the result.
Python library for efficient multi-threaded data processing, with the support for out-of-memory datasets. If you are an R user, chances are that you have already been using the data.table package. Data.table is an extension of the data.frame package in R. It’s also the go-to package for R users when it comes to the fast aggregation of large data (including 1
The relevance of a Doc-string to a function is to guide future users (including future-self), by specifying the right parameters and use cases of the function… So is The Data Science Methodology to data scientists. The Data Science Methodology is an iterative system of methods that guides data scientists on the ideal approach to solving problems with data science, through a prescribed sequence of steps.
OK, welcome to our Word Embedding Series. This post is the first story of the series. You may find this story is suitable for the intermediate or above, who has trained or at least tried once on word2vec, or doc2vec/paragraph vectors. But no worries, I will introduce background, prerequisites and knowledge and how the code implements it from papers in the following posts. I will try my best to do not redirect you to some other links that ask you to read tedious tutorials and end with giving up (trust me, I am the victim of the tremendous online tutorials 🙂 ). I want you to understand word vectors from the coding level together with me so that we can know how to design and implement our word embedding and language model.
There are only a handful of machine learning conferences in the world that attract the top brains in this field. One such conference, which I am an avid follower of, is the International Conference on Machine Learning (ICML). Folks from top machine learning research companies, like Google AI, Facebook, Uber, etc. come together and present their latest research. It’s a conference any data scientist would not want to miss.
Off-Policy Classification – A New Reinforcement Learning Model Selection Method Wednesday, June 19, 2019 Posted by Alex Irpan, Software Engineer, Robotics at Google Reinforcement learning (RL) is a framework that lets agents learn decision making from experience. One of the many variants of RL is off-policy RL, where an agent is trained using a combination of data collected by other agents (off-policy data) and data it collects itself to learn generalizable skills like robotic walking and grasping. In contrast, fully off-policy RL is a variant in which an agent learns entirely from older data, which is appealing because it enables model iteration without requiring a physical robot. With fully off-policy RL, one can train several models on the same fixed dataset collected by previous agents, then select the best one. However, fully off-policy RL comes with a catch: while training can occur without a real robot, evaluation of the models cannot. Furthermore, ground-truth evaluation with a physical robot is too inefficient to test promising approaches that require evaluating a large number of models, such as automated architecture search with AutoML. This challenge motivates off-policy evaluation (OPE), techniques for studying the quality of new agents using data from other agents. With rankings from OPE, we can selectively test only the most promising models on real-world robots, significantly scaling experimentation with the same fixed real robot budget.
Our surveys over the past couple of years have shown growing interest in machine learning (ML) among organizations from diverse industries. A few factors are contributing to this strong interest in implementing ML in products and services. First, the machine learning community has conducted groundbreaking research in many areas of interest to companies, and much of this research has been conducted out in the open via preprints and conference presentations. We are also beginning to see researchers share sample code written in popular open source libraries, and some even share pre-trained models. Organizations now also have more use cases and case studies from which to draw inspiration – no matter what industry or domain you are interested in, chances are there are many interesting ML applications you can learn from. Finally, modeling tools are improving, and automation is beginning to allow new users to tackle problems that used to be the province of experts.
Recently, I had the opportunity to showcase tidymodels in workshops and talks. Because of my vantage point as a user, I figured it would be valuable to share what I have learned so far. Let’s begin by framing where tidymodels fits in our analysis projects. The diagram above is based on the R for Data Science book, by Wickham and Grolemund. The version in this article illustrates what step each package covers. Even though it is a single step, developing models can benefit from having a tidyverse-friendly interface. That is where tidymodels comes in. It is important to clarify that the group of packages that make up tidymodels do not implement statistical models themselves. Instead, they focus on making all the tasks around fitting the model much easier. Those tasks are data pre-processing and results validation.
Training neural networks is tricky. One should be careful that his model is good enough to learn from existing data, and good enough to generalize to unseen data. The lack to generalize a model is mainly because of a problem called overfitting. In simple words, overfitting means that the model achieves very high accuracy on the initial training data and very low accuracy on newly unseen data. It is like when a teacher always gives the same questions in their exams. Their students would easily get very high grades, because they simply memorize the answers. So, the high grades are not a good metric here. A dangerous consequence is that the students will not even bother to learn the subject. There are many techniques to target overfitting, and dropout is one of them. In this article, we will discover what is the intuition behind dropout, how it is used in neural networks, and finally how to implement it in Keras.
In this article I want to explore with you how you can create a containerized data science environment or whatever other systems you might want, that you can quickly deploy to any machine running Docker, may it be your laptop or a cloud computer. The tool I want to demonstrate to you for that purpose is Docker-Compose, an addition to Docker to build and run applications made from multiple containers. The example system I want to build with you in this article will be comprised of three components: a Jupyter Notebook server to conduct experiments in, an MLflow Tracking Server to record and organize experiment parameters and metrics, and a Postgres Database as the backend for the MLflow server and as a handy datastore for your structured datasets. I mostly aim to give you an idea of Docker-Compose and how to use it and will assume that you have at least a basic understanding of Docker or maybe a first idea what it is used for and how it works. If not, let’s take a quick look at why you should bother with yet another technology.
In this article, we’re going to not only train a Convolutional Neural Network using Keras with Python to classify handwritten digits, but also visualize the intermediate activations of the Convolutional Neural Network, in order to gain insights about what features of the image does each layer learn. We will be using the MNIST dataset which can be found here on Kaggle. This dataset contains 42000 rows in the training set and 24000 rows in the testing set. Each row contains 784 pixel values signifying a 28 x 28 image containing handwritten singular digits from 0-9. Let’s dive into the code.
In this article, we discuss the basics of ordinal logistic regression and its implementation in R. Ordinal logistic regression is a widely used classification method, with applications in variety of domains. This method is the go-to tool when there is a natural ordering in the dependent variable. For example, dependent variable with levels low, medium, high is a perfect context for application of logistic ordinal regression.
Whenever I’m developing R code or writing data wrangling or analysis scripts for research projects that I work on I use Emacs and its add-on package Emacs Speaks Statistics (ESS). I’ve done so for nigh on a couple of decades now, ever since I switched full time to running Linux as my daily OS. For years this has served me well, though I wouldn’t call myself an Emacs expert; not even close! With a bit of help from some R Core coding standards document I got indentation working how I like it, I learned to contort my fingers in weird and wonderful ways to execute a small set of useful shortcuts, and I even committed some of those shortcuts to memory. More recently, however, my go-to methods for configuring Emacs+ESS were failing; indentation was all over the shop, the smart _ stopped working or didn’t work as it had for over a decade, syntax highlighting of R-related files, like .Rmd was hit and miss, and polymode was just a mystery to me. Configuring Emacs+ESS was becoming much more of a chore, and rather unhelpfully, my problems coincided with my having less and less time to devote to tinkering with my computer setups. Also, fiddling with this stuff just wasn’t fun any more. So, in a fit of pique following one to many reconfiguration sessions of Emacs+ESS, I went in search of some greener grass. During that search I came across radian, a neat, attractive, simple console for working with R.
Survival analysis is used to analyze the time until the occurrence of an event (or multiple events). Cox models – which are often referred to as semiparametric because they do not assume any particular baseline survival distribution – are perhaps the most widely used technique; however, Cox models are not without limitations and parametric approaches can be advantageous in many contexts. For instance, parametric survival models are essential for extrapolating survival outcomes beyond the available follow-up data. For this reason they are nearly always used in health-economic evaluations where it is necessary to consider the lifetime health effects (and costs) of medical interventions. R contains a large number of packages related to biostatistics and its support for parametric survival modeling is no different. Below we will examine a range of parametric survival distributions, their specifications in R, and the hazard shapes they support. We will then show how the flexsurv package can make parametric regression modeling of survival data straightforward.
This is the second part of an article highlighting some key best practices in the implementation of machine learning solutions. The first part highlighted some of the key differences between the lifecycle of machine learning solutions and traditional software application which are conducive to very unique challenges. Today, I would like to include a few more relevant challenges and solutions that are constantly faced by data science teams when delivering machine learning solutions at a decent scale. To pick up where we left off, let’s continue with the challenges related to data experimentation.
Recently, I’ve been touring different conferences presenting a talk about best practices for implementing large scale machine learning solutions. The idea is to present a series of non-obvious ideas that result incredibly practical in the implementation of machine intelligence applications in the real world. All the lessons have been based on our experiences at Invector Labs working with large organizations and ambitious startups in the implementation of machine learning capabilities. During those exercises, we quickly realized that many of our assumptions of machine learning apps were really flawed and that there was a huge gap between the advancements in AI research and the practical viability of those ideas. In this two-part article, I would like summarize some of those ideas that hopefully will result valuable to machine learning practitioners and aspirational data scientists. There are many challenges that surface in the implementation of real world machine learning solutions. Most of them are related to the mismatch between the lifecycle of machine learning programs and traditional software applications. With some exceptions, traditional software applications follow a relatively sequential model from design to production. Machine learning models, on the other hand, follow a circular lifecycle that include aspects such as regularization or optimization that have no equivalent in the current toolset of traditional software applications.
A statistical test is a tool to make a quantitative decision about sampled data. Although these tests are strongly related to statistical inference, one should keep in his mind that these tests are designed neither to prove nor to falsify the hypothesis. This is the reason there is no such an expression /conclusion in statistics as:
‘There is sufficient evidence at the alpha level of significance to reject the claim given in the alternative hypothesis, so the alternative hypothesis is accepted. ‘
Instead, there are two possible conclusions after employing statistical tests as:
• There is sufficient evidence at the alpha level of significance to reject the claim given in the null hypothesis, so the null hypothesis can be rejected.
• There is no sufficient evidence at the alpha level of significance to reject the claim given in the null hypothesis, so the null hypothesis cannot be rejected.
There is a growing frustration within the data science/machine learning community to see yet another PoC (Proof of Concept) create promising results but never turning into something impactful. The main reason: a large gap between development and deployment. There is an increasing number of tools and frameworks – open source as well as commercial – that promise to bridge this gap. However, introducing them into an organization is a time-consuming and expensive endeavor. Fighting for such solutions is honorable and necessary. Still, there is something else we can do to narrow the gap in the meantime: aligning our workflow with software development principles. Number one on things to learn: unit tests. Unfortunately, most introductions and documentations deal with this topic in an abstract way. In contrast, this blog walks you through a practical workflow with a specific example. Let’s get started!
People are consuming and generating huge volumes of data knowingly and unknowingly on a daily basis. It is this bombardment of digital information is what current businesses are trying to tap and harness to sell and engage their customers more. All types of Industries are bringing a personal touch into their services and offerings to give awesome user experience to their customers. All these have become possible due to powerful Data science enabled AI/ML techniques which are empowering our machines, allowing them to take analytical decisions based on a sea of data accessible to them.
In this tutorial, an end to end project is created in order to do intelligent, realtime and scalable video processing in Azure. In this, a capability is created that can detect graffiti and identify wagon numbers using videos of trains. Properties of the project are as follows:
• Intelligent algorithms to detect graffiti and identify wagon numbers
• Realtime and reliable way of processing videos from edge to cloud
• Scalable for exponential growth of number of videos
• Functional project that can be optimized to any video processing capability
Deep learning image augmentation pipelines typically offer speed or flexibility, but never both at the same time. Computationally efficient, production-ready computer vision pipelines tend to be written in C++ and require developers to specify all the nuts and bolts of image transform algorithms to such an extent that these pipelines end up not terribly amenable to further on-the-fly tweaking. On the other end of the spectrum, popular Python libraries like Pillow offer high-level APIs that let practitioners choose from seemingly unlimited combinations of tweaks that can be applied to a vast repository of image transform algorithms. Unfortunately, this freedom carries with it the cost of a steep drop-off in performance. The DALI Library attempts to give practitioners the best of both worlds. Its image transform algorithms are themselves written in C++ code that squeezes every last drop of performance out of NVIDIA GPU chips, making it possible to perform image transforms in parallel on a per-batch basis, across however many GPUs a user has access to. The C++ source code is wired up to a user-friendly Python API, through which practitioners can define image transform pipelines that play nice with both the PyTorch and TensorFlow frameworks. In an attempt to ascertain whether DALI indeed delivers both the speed and flexibility it advertises, I spent the better part of one week running a series of my own experiments with the library. Spoiler alert: while DALI absolutely brings the speed, flexibility is still somewhat lacking.
Training a neural net is far from being a straightforward task, as the slightest mistake leads to non-optimal results without any warning. Training depends on many factors and parameters and thus require a thoughtful approach. It is known that the beginning of training (i.e., the first few iterations) is very important. When done improperly, you get bad results – sometimes, the network won’t even learn anything at all! For this reason, the way you initialize the weights of the neural network is one of the key factors to good training. The goal of this article is to explain why initialization is impacting and present a different number of ways to implement it efficiently. We will test our approaches against practical examples. The code uses the fastai library (based on pytorch) and lessons from the last fastai MOOC (which, by the way, is really great!). All experiment notebooks are available in this github repository.
Some readers asked for an easy fact-based (no uncertainty or probability) example to accompany my article ‘Never start with a hypothesis’ to show you how setting up the decision context works. Your wish is my command! Let’s play through two scenarios with two different default actions to see how it works. Imagine I’ve just gotten a call from my friend: ‘Shall we go out tonight?’
Centroid Network Traditional clustering algorithms such as K-means rely heavily on the nature of the chosen metric or data representation. To get meaningful clusters, these representations need to be tailored to the downstream task (e.g. cluster photos by object category, cluster faces by identity). Therefore, we frame clustering as a meta-learning task, few-shot clustering, which allows us to specify how to cluster the data at the meta-training level, despite the clustering algorithm itself being unsupervised. We propose Centroid Networks, a simple and efficient few-shot clustering method based on learning representations which are tailored both to the task to solve and to its internal clustering module. We also introduce unsupervised few-shot classification, which is conceptually similar to few-shot clustering, but is strictly harder than supervised* few-shot classification and therefore allows direct comparison with existing supervised few-shot classification methods. On Omniglot and miniImageNet, our method achieves accuracy competitive with popular supervised few-shot classification algorithms, despite using *no labels* from the support set. We also show performance competitive with state-of-the-art learning-to-cluster methods. …
Wasserstein Auto-Encoder (WAE) We propose the Wasserstein Auto-Encoder (WAE)—a new algorithm for building a generative model of the data distribution. WAE minimizes a penalized form of the Wasserstein distance between the model distribution and the target distribution, which leads to a different regularizer than the one used by the Variational Auto-Encoder (VAE). This regularizer encourages the encoded training distribution to match the prior. We compare our algorithm with several other techniques and show that it is a generalization of adversarial auto-encoders (AAE). Our experiments show that WAE shares many of the properties of VAEs (stable training, encoder-decoder architecture, nice latent manifold structure) while generating samples of better quality, as measured by the FID score. …
Battery Reduction Battery reduction is used to select a subset of m variables from an original set of n variables (m < n) that reproduce a large proportion of the variance in the original set of n variables. There are a number of procedures for performing battery reduction analysis. A popular method involves performing a principal components analysis first to select m components, which account for the salient variance in the original data. Gram-Schmidt orthogonal rotations are then performed to determine the m variables that account for the largest proportion of variance. …
Ordered Neurons LSTM (ON-LSTM) Recurrent neural network (RNN) models are widely used for processing sequential data governed by a latent tree structure. Previous work shows that RNN models (especially Long Short-Term Memory (LSTM) based models) could learn to exploit the underlying tree structure. However, its performance consistently lags behind that of tree-based models. This work proposes a new inductive bias Ordered Neurons, which enforces an order of updating frequencies between hidden state neurons. We show that the ordered neurons could explicitly integrate the latent tree structure into recurrent models. To this end, we propose a new RNN unit: ON-LSTM, which achieve good performances on four different tasks: language modeling, unsupervised parsing, targeted syntactic evaluation, and logical inference. …
Interconnectivity of production machines is a key feature of the Industrial Internet of Things (IIoT). This feature allows for many advantages in producing. Configuration and maintenance gets easier, as access to the given production unit is not necessarily coupled to physical presence. Customized production of goods is easily possible, reducing production times and increasing throughput. There are, however, also dangers to the increasing talkativeness of industrial production machines. The more open a system is, the more points of entry for an attacker exist. Furthermore, the amount of data a production site also increases rapidly due to the integrated intelligence and interconnectivity. To keep track of this data in order to detect attacks and errors in the production site, it is necessary to smartly aggregate and evaluate the data. In this paper, we present a new approach for collecting, aggregating and analysing data from different sources and on three different levels of abstraction. Our model is event-centric, considering every occurrence of information inside the system as an event. In the lowest level of abstraction, singular packets are collected, correlated with log-entries and analysed. On the highest level of abstraction, networks are pictured as a connectivity graph, enriched with information about host-based activities. Furthermore, we describe our work in progress of evaluating our aggregation model on two different system settings. In the first scenario, we verify the usability of our model in a remote maintenance application. In the second scenario, we evaluate our model in the context of network sniffing and correlation with log-files. First results show that our model is a promising solution to cope with increasing amounts of data and to correlate information from different types of sources.
This paper introduces the Contextual Evaluation Model (CEM), a novel method for knowledge representation and manipulation. The CEM differs from existing models in that it integrates facts, patterns and sequences into a single contextual framework. V5, an implementation of the model is presented and demonstrated with multiple annotated examples. The paper includes simulations demonstrating how the model reacts to pleasure/pain stimuli. The ‘thought’ is defined within the model and examples are given converting thoughts to language, converting language to thoughts and how ‘meaning’ arises from thoughts. A pattern learning algorithm is described. The algorithm is applied to multiple problems ranging from recognizing a voice to the autonomous learning of a simplified natural language.
This is a detailed tutorial paper which explains the Principal Component Analysis (PCA), Supervised PCA (SPCA), kernel PCA, and kernel SPCA. We start with projection, PCA with eigen-decomposition, PCA with one and multiple projection directions, properties of the projection matrix, reconstruction error minimization, and we connect to auto-encoder. Then, PCA with singular value decomposition, dual PCA, and kernel PCA are covered. SPCA using both scoring and Hilbert-Schmidt independence criterion are explained. Kernel SPCA using both direct and dual approaches are then introduced. We cover all cases of projection and reconstruction of training and out-of-sample data. Finally, some simulations are provided on Frey and AT&T face datasets for verifying the theory in practice.
We propose a new sequential monitoring scheme for changes in the parameters of a multivariate time series. In contrast to procedures proposed in the literature which compare an estimator from the training sample with an estimator calculated from the remaining data, we suggest to divide the sample at each time point after the training sample. Estimators from the sample before and after all separation points are then continuously compared calculating a maximum of norms of their differences. For open-end scenarios our approach yields an asymptotic level procedure, which is consistent under the alternative of a change in the parameter.
Centrality of emotion for the stories told by humans is underpinned by numerous studies in literature and psychology. The research in automatic storytelling has recently turned towards emotional storytelling, in which characters’ emotions play an important role in the plot development. However, these studies mainly use emotion to generate propositional statements in the form ‘A feels affection towards B’ or ‘A confronts B’. At the same time, emotional behavior does not boil down to such propositional descriptions, as humans display complex and highly variable patterns in communicating their emotions, both verbally and non-verbally. In this paper, we analyze how emotions are expressed non-verbally in a corpus of fan fiction short stories. Our analysis shows that stories written by humans convey character emotions along various non-verbal channels. We find that some non-verbal channels, such as facial expressions and voice characteristics of the characters, are more strongly associated with joy, while gestures and body postures are more likely to occur with trust. Based on our analysis, we argue that automatic storytelling systems should take variability of emotion into account when generating descriptions of characters’ emotions.
Neural Architecture Search (NAS) has been widely studied for designing discriminative deep learning models such as image classification, object detection, and semantic segmentation. As a large number of priors have been obtained through the manual design of architectures in the fields, NAS is usually considered as a supplement approach. In this paper, we have significantly expanded the application areas of NAS by performing an empirical study of NAS to search generative models, or specifically, auto-encoder based universal style transfer, which lacks systematic exploration, if any, from the architecture search aspect. In our work, we first designed a search space where common operators for image style transfer such as VGG-based encoders, whitening and coloring transforms (WCT), convolution kernels, instance normalization operators, and skip connections were searched in a combinatorial approach. With a simple yet effective parallel evolutionary NAS algorithm with multiple objectives, we derived the first group of end-to-end deep networks for universal photorealistic style transfer. Comparing to random search, a NAS method that is gaining popularity recently, we demonstrated that carefully designed search strategy leads to much better architecture design. Finally compared to existing universal style transfer networks for photorealistic rendering such as PhotoWCT that stacks multiple well-trained auto-encoders and WCT transforms in a non-end-to-end manner, the architectures designed by StyleNAS produce better style-transferred images with details preserving, using a tiny number of operators/parameters, and enjoying around 500x inference time speed-up.
In this proceeding we give an overview of the idea of covariance (or equivariance) featured in the recent development of convolutional neural networks (CNNs). We study the similarities and differences between the use of covariance in theoretical physics and in the CNN context. Additionally, we demonstrate that the simple assumption of covariance, together with the required properties of locality, linearity and weight sharing, is sufficient to uniquely determine the form of the convolution.
Bayesian methods promise to fix many shortcomings of deep learning, but they are impractical and rarely match the performance of standard methods, let alone improve them. In this paper, we demonstrate practical training of deep networks with natural-gradient variational inference. By applying techniques such as batch normalisation, data augmentation, and distributed training, we achieve similar performance in about the same number of epochs as the Adam optimiser, even on large datasets such as ImageNet. Importantly, the benefits of Bayesian principles are preserved: predictive probabilities are well-calibrated and uncertainties on out-of-distribution data are improved. This work enables practical deep learning while preserving benefits of Bayesian principles. A PyTorch implementation will be available as a plug-and-play optimiser.
Automatic question generation (QG) is a challenging problem in natural language understanding. QG systems are typically built assuming access to a large number of training instances where each instance is a question and its corresponding answer. For a new language, such training instances are hard to obtain making the QG problem even more challenging. Using this as our motivation, we study the reuse of an available large QG dataset in a secondary language (e.g. English) to learn a QG model for a primary language (e.g. Hindi) of interest. For the primary language, we assume access to a large amount of monolingual text but only a small QG dataset. We propose a cross-lingual QG model which uses the following training regime: (i) Unsupervised pretraining of language models in both primary and secondary languages and (ii) joint supervised training for QG in both languages. We demonstrate the efficacy of our proposed approach using two different primary languages, Hindi and Chinese. We also create and release a new question answering dataset for Hindi consisting of 6555 sentences.
Rain removal has recently attracted increasing research attention, as it is able to enhance the visibility of rain videos. However, the existing learning based rain removal approaches for videos suffer from insufficient training data, especially when applying deep learning to remove rain. In this paper, we establish a large-scale video database for rain removal (LasVR), which consists of 316 rain videos. Then, we observe from our database that there exist the temporal correlation of clean content and similar patterns of rain across video frames. According to these two observations, we propose a two-stream convolutional long- and short- term memory (ConvLSTM) approach for rain removal in videos. The first stream is composed of the subnet for rain detection, while the second stream is the subnet of rain removal that leverages the features from the rain detection subnet. Finally, the experimental results on both synthetic and real rain videos show the proposed approach performs better than other state-of-the-art approaches.
Modern machine learning methods including deep learning have achieved great success in predictive accuracy for supervised learning tasks, but may still fall short in giving useful estimates of their predictive {\em uncertainty}. Quantifying uncertainty is especially critical in real-world settings, which often involve input distributions that are shifted from the training distribution due to a variety of factors including sample bias and non-stationarity. In such settings, well calibrated uncertainty estimates convey information about when a model’s output should (or should not) be trusted. Many probabilistic deep learning methods, including Bayesian-and non-Bayesian methods, have been proposed in the literature for quantifying predictive uncertainty, but to our knowledge there has not previously been a rigorous large-scale empirical comparison of these methods under dataset shift. We present a large-scale benchmark of existing state-of-the-art methods on classification problems and investigate the effect of dataset shift on accuracy and calibration. We find that traditional post-hoc calibration does indeed fall short, as do several other previous methods. However, some methods that marginalize over models give surprisingly strong results across a broad spectrum of tasks.
Contextual information, such as the co-occurrence of objects and the spatial and relative size among objects provides deep and complex information about scenes. It also can play an important role in improving object detection. In this work, we present two contextual models (rescoring and re-labeling models) that leverage contextual information (16 contextual relationships are applied in this paper) to enhance the state-of-the-art RCNN-based object detection (Faster RCNN). We experimentally demonstrate that our models lead to enhancement in detection performance using the most common dataset used in this field (MSCOCO).
DNN pruning reduces memory footprint and computational work of DNN-based solutions to improve performance and energy-efficiency. An effective pruning scheme should be able to systematically remove connections and/or neurons that are unnecessary or redundant, reducing the DNN size without any loss in accuracy. In this paper we show that prior pruning schemes require an extremely time-consuming iterative process that requires retraining the DNN many times to tune the pruning hyperparameters. We propose a DNN pruning scheme based on Principal Component Analysis and relative importance of each neuron’s connection that automatically finds the optimized DNN in one shot without requiring hand-tuning of multiple parameters.
In social network Twitter, users can interact with each other and spread information via retweets. These millions of interactions may result in media events whose influence goes beyond Twitter framework. In this paper, we thoroughly explore interactions to provide a better understanding of the emergence of certain trends. First, we consider an interaction on Twitter to be a triplet meaning that user , called the spreader, has retweeted a tweet of user , called the author, at time . We model this set of interactions as a data cube with three dimensions: spreaders, authors and time. Then, we provide a method which builds different contexts, where a context is a set of features characterizing the circumstances of an event. Finally, these contexts allow us to find relevant unexpected behaviors, according to several dimensions and various perspectives: a user during a given hour which is abnormal compared to its usual behavior, a relationship between two users which is abnormal compared to all other relationships, \textit{etc.} We apply our method to a set of retweets related to the 2017 French presidential election and show that one can build interesting insights regarding political organization on Twitter.
Lebesgue sampling is based on collecting information depending on the values of the signal. Although the interpolation methods for periodic sampling have been a topic of research for a long time, there is a lack of study in methods capable of taking advantage of the Lebesgue sampling characteristics to reconstruct time series more accurately. Indeed, Lebesgue sampling contains additional information about the shape of the signal in-between two sampled points. Using this information would allow us to generate an interpolated signal closer to the original one. That is to say, the average distance between the interpolated signal and the original signal will be smaller than a signal interpolated with other interpolation methods. In this paper, we propose two novel time series interpolation methods specifically designed for Lebesgue sampling called ZeLiC and ZeChipC. ZeLiC is an algorithm that combines both Zero-order hold interpolation and Linear interpolation to reconstruct time series. ZeChipC is a similar idea, it is a combination of Zero-order hold and PCHIP interpolation. Zero-order hold interpolation is favourable for interpolating abrupt changes while Linear and PCHIP interpolation are more suitable for smooth transitions. In order to apply one method or the other, we have introduced a new concept called tolerated region. ZeLiC and ZeChipC include a new functionality to adapt the reconstructed signal to concave/convex regions. The proposed methods have been compared with the state-of-the-art interpolation methods using Lebesgue sampling and have offered higher average performance. Additionally, we have compared the performance of the methods using both Riemann and Lebesgue sampling using an approximate number of sampled points. The performance of the combination ‘Lebesgue sampling with ZeChipC interpolation method’ is clearly much better than any other combination.
A graphical model is a structured representation of the data generating process. The traditional method to reason over random variables is to perform inference in this graphical model. However, in many cases the generating process is only a poor approximation of the much more complex true data generating process, leading to suboptimal estimation. The subtleties of the generative process are however captured in the data itself and we can `learn to infer’, that is, learn a direct mapping from observations to explanatory latent variables. In this work we propose a hybrid model that combines graphical inference with a learned inverse model, which we structure as in a graph neural network, while the iterative algorithm as a whole is formulated as a recurrent neural network. By using cross-validation we can automatically balance the amount of work performed by graphical inference versus learned inference. We apply our ideas to the Kalman filter, a Gaussian hidden Markov model for time sequences, and show, among other things, that our model can estimate the trajectory of a noisy chaotic Lorenz Attractor much more accurately than either the learned or graphical inference run in isolation.
Cost and cardinality estimation is vital to query optimizer, which can guide the plan selection. However traditional empirical cost and cardinality estimation techniques cannot provide high-quality estimation, because they cannot capture the correlation between multiple columns. Recently the database community shows that the learning-based cardinality estimation is better than the empirical methods. However, existing learning-based methods have several limitations. Firstly, they can only estimate the cardinality, but cannot estimate the cost. Secondly, convolutional neural network (CNN) with average pooling is hard to represent complicated structures, e.g., complex predicates, and the model is hard to be generalized. To address these challenges, we propose an effective end-to-end learning-based cost estimation framework based on a tree-structured model, which can estimate both cost and cardinality simultaneously. To the best of our knowledge, this is the first end-to-end cost estimator based on deep learning. We propose effective feature extraction and encoding techniques, which consider both queries and physical operations in feature extraction. We embed these features into our tree-structured model. We propose an effective method to encode string values, which can improve the generalization ability for predicate matching. As it is prohibitively expensive to enumerate all string values, we design a patten-based method, which selects patterns to cover string values and utilizes the patterns to embed string values. We conducted experiments on real-world datasets and experimental results showed that our method outperformed baselines.
Artificial neural networks (ANNs) suffer from catastrophic forgetting when trained on a sequence of tasks. While this phenomenon was studied in the past, there is only very limited recent research on this phenomenon. We propose a method for determining the contribution of individual parameters in an ANN to catastrophic forgetting. The method is used to analyze an ANNs response to three different continual learning scenarios.
Accessibility is a major challenge of machine learning (ML). Typical ML models are built by specialists and require specialized hardware/software as well as ML experience to validate. This makes it challenging for non-technical collaborators and endpoint users (e.g. physicians) to easily provide feedback on model development and to gain trust in ML. The accessibility challenge also makes collaboration more difficult and limits the ML researcher’s exposure to realistic data and scenarios that occur in the wild. To improve accessibility and facilitate collaboration, we developed an open-source Python package, Gradio, which allows researchers to rapidly generate a visual interface for their ML models. Gradio makes accessing any ML model as easy as sharing a URL. Our development of Gradio is informed by interviews with a number of machine learning researchers who participate in interdisciplinary collaborations. Their feedback identified that Gradio should support a variety of interfaces and frameworks, allow for easy sharing of the interface, allow for input manipulation and interactive inference by the domain expert, as well as allow embedding the interface in iPython notebooks. We developed these features and carried out a case study to understand Gradio’s usefulness and usability in the setting of a machine learning collaboration between a researcher and a cardiologist.
The Maximum k-plex Problem is an important combinatorial optimization problem with increasingly wide applications. Due to its exponential time complexity, many heuristic methods have been proposed which can return a good-quality solution in a reasonable time. However, most of the heuristic algorithms are memoryless and unable to utilize the experience during the search. Inspired by the multi-armed bandit (MAB) problem in reinforcement learning (RL), we propose a novel perturbation mechanism named BLP, which can learn online to select a good vertex for perturbation when getting stuck in local optima. To our best of knowledge, this is the first attempt to combine local search with RL for the maximum -plex problem. Besides, we also propose a novel strategy, named Dynamic-threshold Configuration Checking (DTCC), which extends the original Configuration Checking (CC) strategy from two aspects. Based on the BLP and DTCC, we develop a local search algorithm named BDCC and improve it by a hyperheuristic strategy. The experimental result shows that our algorithms dominate on the standard DIMACS and BHOSLIB benchmarks and achieve state-of-the-art performance on massive graphs.
This work presents a suboptimality study of a particular model predictive control with a stage cost shaping based on the ideas of reinforcement learning. The focus of the suboptimality study is to derive quantities relating the infinite-horizon cost function under the said variant of model predictive control to the respective infinite-horizon value function. The basis control scheme involves usual stabilizing constraints comprising of a terminal set and a terminal cost in the form of a local Lyapunov function. The stage cost is adapted using the principles of Q-learning, a particular approach to reinforcement learning. The work is concluded by case studies with two systems for wide ranges of initial conditions.
This paper proposes Quaternion Collaborative Filtering (QCF), a novel representation learning method for recommendation. Our proposed QCF relies on and exploits computation with Quaternion algebra, benefiting from the expressiveness and rich representation learning capability of Hamilton products. Quaternion representations, based on hypercomplex numbers, enable rich inter-latent dependencies between imaginary components. This encourages intricate relations to be captured when learning user-item interactions, serving as a strong inductive bias as compared with the real-space inner product. All in all, we conduct extensive experiments on six real-world datasets, demonstrating the effectiveness of Quaternion algebra in recommender systems. The results exhibit that QCF outperforms a wide spectrum of strong neural baselines on all datasets. Ablative experiments confirm the effectiveness of Hamilton-based composition over multi-embedding composition in real space.
We introduce multi-frequency vector diffusion maps (MFVDM), a new framework for organizing and analyzing high dimensional datasets. The new method is a mathematical and algorithmic generalization of vector diffusion maps (VDM) and other non-linear dimensionality reduction methods. MFVDM combines different nonlinear embeddings of the data points defined with multiple unitary irreducible representations of the alignment group that connect two nodes in the graph. We illustrate the efficacy of MFVDM on synthetic data generated according to a random graph model and cryo-electron microscopy image dataset. The new method achieves better nearest neighbor search and alignment estimation than the state-of-the-arts VDM and diffusion maps (DM) on extremely noisy data.
The process of knowledge acquisition can be viewed as a question-answer game between a student and a teacher in which the student typically starts by asking broad, open-ended questions before drilling down into specifics (Hintikka, 1981; Hakkarainen and Sintonen, 2002). This pedagogical perspective motivates a new way of representing documents. In this paper, we present SQUASH (Specificity-controlled Question-Answer Hierarchies), a novel and challenging text generation task that converts an input document into a hierarchy of question-answer pairs. Users can click on high-level questions (e.g., ‘Why did Frodo leave the Fellowship?’) to reveal related but more specific questions (e.g., ‘Who did Frodo leave with?’). Using a question taxonomy loosely based on Lehnert (1978), we classify questions in existing reading comprehension datasets as either ‘general’ or ‘specific’. We then use these labels as input to a pipelined system centered around a conditional neural language model. We extensively evaluate the quality of the generated QA hierarchies through crowdsourced experiments and report strong empirical results.
In this work we study loss functions for learning and evaluating probability distributions over large discrete domains. Unlike classification or regression where a wide variety of loss functions are used, in the distribution learning and density estimation literature, very few losses outside the dominant are applied. We aim to understand this fact, taking an axiomatic approach to the design of loss functions for learning distributions. We start by proposing a set of desirable criteria that any good loss function should satisfy. Intuitively, these criteria require that the loss function faithfully evaluates a candidate distribution, both in expectation and when estimated on a few samples. Interestingly, we observe that \emph{no loss function} possesses all of these criteria. However, one can circumvent this issue by introducing a natural restriction on the set of candidate distributions. Specifically, we require that candidates are with respect to the target distribution, i.e., they may contain less information than the target but otherwise do not significantly distort the truth. We show that, after restricting to this set of distributions, the log loss, along with a large variety of other losses satisfy the desired criteria. These results pave the way for future investigations of distribution learning that look beyond the log loss, choosing a loss function based on application or domain need.
We propose a new objective function for finite-horizon episodic Markov decision processes that better captures Bellman’s principle of optimality, and provide an expression for the gradient of the objective.
Deep approaches to anomaly detection have recently shown promising results over shallow approaches on high-dimensional data. Typically anomaly detection is treated as an unsupervised learning problem. In practice however, one may have—in addition to a large set of unlabeled samples—access to a small pool of labeled samples, e.g. a subset verified by some domain expert as being normal or anomalous. Semi-supervised approaches to anomaly detection make use of such labeled data to improve detection performance. Few deep semi-supervised approaches to anomaly detection have been proposed so far and those that exist are domain-specific. In this work, we present Deep SAD, an end-to-end methodology for deep semi-supervised anomaly detection. Using an information-theoretic perspective on anomaly detection, we derive a loss motivated by the idea that the entropy for the latent distribution of normal data should be lower than the entropy of the anomalous distribution. We demonstrate in extensive experiments on MNIST, Fashion-MNIST, and CIFAR-10 along with other anomaly detection benchmark datasets that our approach is on par or outperforms shallow, hybrid, and deep competitors, even when provided with only few labeled training data.
Transformer architectures show significant promise for natural language processing. Given that a single pretrained model can be fine-tuned to perform well on many different tasks, these networks appear to extract generally useful linguistic features. A natural question is how such networks represent this information internally. This paper describes qualitative and quantitative investigations of one particularly effective model, BERT. At a high level, linguistic features seem to be represented in separate semantic and syntactic subspaces. We find evidence of a fine-grained geometric representation of word senses. We also present empirical descriptions of syntactic representations in both attention matrices and individual word embeddings, as well as a mathematical argument to explain the geometry of these representations.
Planning and Learning are complementary approaches. Planning relies on deliberative reasoning about the current state and sequence of future reachable states to solve the problem. Learning, on the other hand, is focused on improving system performance based on experience or available data. Learning to improve the performance of planning based on experience in similar, previously solved problems, is ongoing research. One approach is to learn Value function (cost-to-go) which can be used as heuristics for speeding up search-based planning. Existing approaches in this direction use the results of the previous search for learning the heuristics. In this work, we present a search-inspired approach of systematic model exploration for the learning of the value function which does not stop when a plan is available but rather prolongs search such that not only resulting optimal path is used but also extended region around the optimal path. This, in turn, improves both the efficiency and robustness of successive planning. Additionally, the effect of losing admissibility by using ML heuristic is managed by bounding ML with other admissible heuristics.
We introduce a pooling method for sets of feature vectors based on sorting features across elements of the set. This allows a deep neural network for sets to learn more flexible representations. We also demonstrate how FSPool can be used to construct a permutation-equivariant auto-encoder. On a toy dataset of polygons and a set version of MNIST, we show that such an auto-encoder produces considerably better reconstructions. Used in set classification, FSPool significantly improves accuracy and convergence speed on the set versions of MNIST and CLEVR.
A novel semi-supervised learning technique is introduced based on a simple iterative learning cycle together with learned thresholding techniques and an ensemble decision support system. State-of-the-art model performance and increased training data volume are demonstrated, through the use of unlabelled data when training deeply learned classification models. Evaluation of the proposed approach is performed on commonly used datasets when evaluating semi-supervised learning techniques as well as a number of more challenging image classification datasets (CIFAR-100 and a 200 class subset of ImageNet).
Neural architecture search (NAS), or automated design of neural network models, remains a very challenging meta-learning problem. Several recent works (called ‘one-shot’ approaches) have focused on dramatically reducing NAS running time by leveraging proxy models that still provide architectures with competitive performance. In our work, we propose a new meta-learning algorithm that we call CoNAS, or Compressive sensing-based Neural Architecture Search. Our approach merges ideas from one-shot approaches with iterative techniques for learning low-degree sparse Boolean polynomial functions. We validate our approach on several standard test datasets, discover novel architectures hitherto unreported, and achieve competitive (or better) results in both performance and search time compared to existing NAS approaches. Further, we support our algorithm with a theoretical analysis, providing upper bounds on the number of measurements needed to perform reliable meta-learning; to our knowledge, these analysis tools are novel to the NAS literature and may be of independent interest.
-NN classifier. The results characterize precisely the contribution of the observations from the source distribution to the classification task under the target distribution. A data-driven adaptive classifier is then proposed and is shown to simultaneously attain within a logarithmic factor of the optimal rate over a large collection of parameter spaces. Simulation studies and real data applications are carried out where the numerical results further illustrate the theoretical analysis. Extensions to the case of multiple source distributions are also considered.
partitions and 
learning rate, where 
is the labeled sample size. Further, we derive more refined results by using additional unlabeled data to enlarge the number of partitions and by generating features in a data-dependent way to reduce the number of random features.
procedure, which is consistent under the alternative of a change in the parameter.
meaning that user 
, called the spreader, has retweeted a tweet of user 
, called the author, at time 
. We model this set of interactions as a data cube with three dimensions: spreaders, authors and time. Then, we provide a method which builds different contexts, where a context is a set of features characterizing the circumstances of an event. Finally, these contexts allow us to find relevant unexpected behaviors, according to several dimensions and various perspectives: a user during a given hour which is abnormal compared to its usual behavior, a relationship between two users which is abnormal compared to all other relationships, \textit{etc.} We apply our method to a set of retweets related to the 2017 French presidential election and show that one can build interesting insights regarding political organization on Twitter.
-plex problem. Besides, we also propose a novel strategy, named Dynamic-threshold Configuration Checking (DTCC), which extends the original Configuration Checking (CC) strategy from two aspects. Based on the BLP and DTCC, we develop a local search algorithm named BDCC and improve it by a hyperheuristic strategy. The experimental result shows that our algorithms dominate on the standard DIMACS and BHOSLIB benchmarks and achieve state-of-the-art performance on massive graphs.
are applied. We aim to understand this fact, taking an axiomatic approach to the design of loss functions for learning distributions. We start by proposing a set of desirable criteria that any good loss function should satisfy. Intuitively, these criteria require that the loss function faithfully evaluates a candidate distribution, both in expectation and when estimated on a few samples. Interestingly, we observe that \emph{no loss function} possesses all of these criteria. However, one can circumvent this issue by introducing a natural restriction on the set of candidate distributions. Specifically, we require that candidates are 
with respect to the target distribution, i.e., they may contain less information than the target but otherwise do not significantly distort the truth. We show that, after restricting to this set of distributions, the log loss, along with a large variety of other losses satisfy the desired criteria. These results pave the way for future investigations of distribution learning that look beyond the log loss, choosing a loss function based on application or domain need.